Introduction
Explainable AI (XAI) has traditionally focused on providing "instant explanations"—immediate, low-effort insights like heatmaps or feature weights. However, these snapshots often fail to support users in complex, high-stakes environments where understanding a model's long-term behavior and knowing when to trust it are critical. This paper argues for a paradigm shift: XAI should move from one-way information delivery to supporting robust, long-term human learning. By integrating principles from the learning sciences, such as "backward design" and "deliberate practice," we can help users build more accurate mental models of AI systems.
The Learner-Centered XAI Framework
The proposed framework is built on three core pillars derived from the learning sciences:
- Backward Design: Rather than starting with a specific XAI technique, designers begin by identifying desired learning objectives (what should the user be able to do?). They then define measures to assess learning and finally design learning activities (the XAI interface) to meet those goals.
- Participatory Design for Learning: This involves co-designing learning objectives with stakeholders (e.g., domain experts and end-users) to ensure the XAI addresses real-world "value tensions" and practical needs.
- Closing the Loop: An iterative process where the effectiveness of the XAI is evaluated based on actual learning outcomes, allowing for continuous refinement of the system.
Case Study: Child Maltreatment Screening
The framework was applied to the Allegheny Family Screening Tool (AFST), an algorithmic decision support tool used in Pennsylvania for child maltreatment screening. In this context, a significant gap exists: the AI predicts long-term risk, while social workers focus on immediate safety. A learner-centered approach helps identify these misalignments, designing XAI that assists workers in integrating these distinct perspectives rather than simply mimicking the AI's output.
Deliberate Practice and Interface Prototypes
A key insight is treating XAI as a form of deliberate practice. Some AI concepts are better grasped through sustained engagement rather than static explanations. For example, "score guessing" exercises can help users mentally simulate a model's behavior more effectively than a feature importance chart.
To demonstrate this, two prototype training interfaces were developed using real historical data from the AFST context:
- Simulated Decision-Making: Users practice making decisions in a low-stakes environment and receive immediate feedback to help calibrate their reliance on the model.
- Iterative Score Guessing: Users attempt to predict the model's risk score, building a more nuanced mental model of the AI's underlying logic through trial and error.
Conclusion and Future Directions
By forcing a "backward shift" in design—prioritizing learning objectives and measurements over specific technical implementations—this framework prevents the creation of "explainable" tools that fail to improve human performance. While initial results in the social work domain are promising, future work will involve full-scale field evaluations and testing the framework's generalizability in other high-stakes domains like content moderation.
Summary updated